brainy/docs/transactions.md

---
title: Transactions & Atomicity
slug: guides/transactions
public: true
category: guides
template: guide
order: 10
description: How Brainy keeps every write atomic — automatic per-operation transactions with rollback, and brain.transact() for atomic multi-write batches with compare-and-swap.
next:
  - concepts/consistency-model
  - guides/optimistic-concurrency
---

# Transaction System

**Status:** ✅ Production Ready

## Overview

Brainy's transaction system provides **atomic operations** with automatic rollback on failure. All operations within a transaction either succeed completely or fail completely - there are no partial failures.

### Key Benefits

- **Atomicity**: All operations succeed or all rollback
- **Consistency**: Indexes and storage remain consistent
- **Automatic**: Transparently used by all `brain.add()`, `brain.update()`, `brain.delete()`, and `brain.relate()` operations
- **Composable**: `brain.transact()` runs a multi-write batch through the same machinery as exactly one atomic commit

## Architecture

```
User Code (brain.add(), brain.update(), brain.transact(), etc.)
 ↓
Transaction Manager (orchestration)
 ↓
Operations (SaveNounMetadataOperation, SaveNounOperation, etc.)
 ↓
Storage Adapter (sharding, ID-first routing)
```

### How It Works

Every write operation in Brainy automatically uses transactions:

```typescript
// Internally, this uses a transaction
const id = await brain.add({
 data: { name: 'Alice', role: 'Engineer' },
 type: NounType.Person
})
```

**Transaction Flow:**

1. **Begin Transaction**: TransactionManager creates new transaction
2. **Add Operations**: Operations added to transaction (SaveNounMetadataOperation, SaveNounOperation)
3. **Execute**: Each operation executes in sequence
4. **Commit**: All operations succeeded → changes persist
5. **Rollback**: Any operation failed → all changes reverted

### Rollback Mechanism

Each operation implements both **execute** and **undo**:

```typescript
class SaveNounMetadataOperation {
 async execute(): Promise<void> {
 // Save new metadata
 await this.storage.saveNounMetadata(this.id, this.metadata)
 }

 async undo(): Promise<void> {
 // Restore previous metadata (or delete if new entity)
 if (this.previousMetadata) {
 await this.storage.saveNounMetadata(this.id, this.previousMetadata)
 } else {
 await this.storage.deleteNounMetadata(this.id)
 }
 }
}
```

**On failure:**
- Operations rolled back in **reverse order**
- Previous state fully restored
- Indexes updated to reflect rollback

## Compatibility with Advanced Features

### Multi-Write Batches: `brain.transact()`

✅ **The 8.0 path for atomic multi-entity writes**

Single-operation methods each commit their own transaction. When several
writes must succeed or fail **together**, use `brain.transact()` — a
declarative batch that commits as exactly one generation, with optional
whole-store compare-and-swap and durable transaction metadata:

```typescript
const db = await brain.transact([
 { op: 'add', id: orderId, type: NounType.Document, subtype: 'order', data: 'Order #1042' },
 { op: 'update', id: customerId, metadata: { lastOrderAt: Date.now() }, ifRev: customer._rev },
 { op: 'relate', from: customerId, to: orderId, type: VerbType.Creates, subtype: 'purchase' }
], { meta: { author: 'order-service' } })

db.receipt.ids // resolved id per operation, in input order
```

**How It Works:**
- The batch executes through the same TransactionManager as single
 operations, wrapped in the generational commit protocol: before-images are
 staged and fsynced first, and the atomic manifest rename is the commit
 point — a crash anywhere before it rolls back to the exact
 pre-transaction bytes.
- Per-entity `ifRev` and whole-store `ifAtGeneration` provide
 compare-and-swap at two granularities; any conflict rejects the entire
 batch before anything is staged.
- The returned `Db` is a pinned, snapshot-isolated view of the committed
 state.

See the **[consistency model](concepts/consistency-model.md)** for the
full guarantees (snapshot isolation, time travel, snapshots) and
**[Snapshots & Time Travel](guides/snapshots-and-time-travel.md)** for
recipes.

### Sharding

✅ **Fully Compatible**

Transactions work across multiple shards:

```typescript
// Entities with different UUID prefixes go to different shards
const id1 = 'aaa00000-1111-4111-8111-111111111111' // Shard: aaa
const id2 = 'bbb00000-2222-4222-8222-222222222222' // Shard: bbb

await brain.add({ id: id1, data: { name: 'Entity A' }, type: NounType.Thing })
await brain.relate({ from: id1, to: id2, type: VerbType.RelatesTo })

// Transaction handles cross-shard atomicity automatically
```

**How It Works:**
- Sharding is transparent to transactions
- `analyzeKey()` method routes to correct shard based on UUID
- Transaction operations don't need to know about shards
- Rollback works across all shards involved

### ID-First Storage

✅ **Fully Compatible**

Transactions work with direct ID-first paths - no type routing needed!

```typescript
// Entities stored with direct ID-first paths
const personId = await brain.add({
 data: { name: 'John Doe' },
 type: NounType.Person // → entities/nouns/{shard}/{id}/metadata.json
})

const orgId = await brain.add({
 data: { name: 'Acme Corp' },
 type: NounType.Organization // → entities/nouns/{shard}/{id}/metadata.json
})

// Type changes handled atomically (type is just metadata)
await brain.update({
 id: personId,
 type: NounType.Organization, // Type change
 data: { name: 'Doe Corp' }
})
```

**How It Works:**
- Type information stored in metadata.noun field
- Storage layer uses O(1) ID-first path construction
- No type cache needed (removed in a previous version)
- Type counters adjusted on commit/rollback
- 40x faster path lookups (eliminates 42-type search)

### Storage Adapter Interface

✅ **Fully Compatible**

Transactions go through the `StorageAdapter` interface, so both shipped adapters (filesystem, memory) and any custom plugin adapter inherit the same atomicity guarantees:

```typescript
const brain = new Brainy({
  storage: { type: 'filesystem', rootDirectory: './data' }
})

await brain.add({ data: { name: 'Entity' }, type: NounType.Thing })
```

**How It Works:**
- Transactions operate through `StorageAdapter` interface
- Custom adapters registered via the plugin system implement the same interface
- Atomicity guaranteed at the write-coordinator level
- Read-after-write consistency maintained inside a single Brainy process

## Examples

### Basic Add Operation

```typescript
import { Brainy } from '@soulcraft/brainy'
import { NounType } from '@soulcraft/brainy/types'

const brain = new Brainy()
await brain.init()

// Automatically uses transaction
const id = await brain.add({
 data: { name: 'Alice', role: 'Engineer' },
 type: NounType.Person
})

// If add fails, all changes rolled back automatically
```

### Update with Type Change

```typescript
// Original entity
const id = await brain.add({
 data: { name: 'John Smith', category: 'individual' },
 type: NounType.Person
})

// Update with type change (atomic)
await brain.update({
 id,
 type: NounType.Organization, // Type change
 data: { name: 'Smith Corp', category: 'business' }
})

// If update fails, original type and data restored
```

### Creating Relationships

```typescript
const personId = await brain.add({
 data: { name: 'Alice' },
 type: NounType.Person
})

const projectId = await brain.add({
 data: { name: 'Project X' },
 type: NounType.Thing
})

// Create relationship (atomic)
await brain.relate({
 from: personId,
 to: projectId,
 type: VerbType.WorksOn
})

// If relate fails, no partial relationship created
```

### Batch Operations

```typescript
// Multiple operations, all atomic
for (let i = 0; i < 100; i++) {
 await brain.add({
 data: { name: `Entity ${i}`, index: i },
 type: NounType.Thing
 })
}

// Each add() is a separate transaction
// If any add fails, only that specific add is rolled back
```

### Delete with Cascade

```typescript
const personId = await brain.add({
 data: { name: 'Bob' },
 type: NounType.Person
})

const projectId = await brain.add({
 data: { name: 'Project Y' },
 type: NounType.Thing
})

await brain.relate({
 from: personId,
 to: projectId,
 type: VerbType.WorksOn
})

// Delete person (atomic - deletes entity + relationships)
await brain.delete(personId)

// If delete fails, both entity and relationships remain
```

## Error Handling

Transactions automatically handle errors and rollback:

```typescript
try {
 await brain.add({
 data: { name: 'Test Entity' },
 type: NounType.Thing,
 vector: [1, 2, 3] // Wrong dimension → error
 })
} catch (error) {
 // Transaction automatically rolled back
 // No partial data in storage or indexes
 console.error('Add failed:', error.message)
}
```

**Common Error Scenarios:**
- **Invalid vector dimension**: Automatic rollback
- **Type validation failure**: Automatic rollback
- **Storage write failure**: Automatic rollback
- **Index update failure**: Automatic rollback

## Performance Considerations

### Transaction Overhead

**What a transaction costs:**
- A typical single-operation write wraps 2-8 operations (metadata + data + indexes) in one transaction
- The overhead is bookkeeping (operation objects + undo state), not extra I/O on the success path
- Rollback cost is proportional to the operations already applied (each is undone in reverse order)

**Optimization:**
- Operations executed sequentially (not parallel) for consistency
- Rollback only happens on failure (success path is fast)
- Index updates batched within transaction

### Auditing Committed Batches

Every committed `brain.transact()` batch is recorded in the transaction
log, newest first:

```typescript
await brain.transact(ops, { meta: { author: 'import-job' } })

const entries = await brain.transactionLog({ limit: 10 })
// [{ generation: 1042, timestamp: 1765432100000, meta: { author: 'import-job' } }]
```

Single-operation writes advance the generation counter but do not append
log entries — see the [consistency model](concepts/consistency-model.md)
for the history-granularity contract.

## Best Practices

### 1. Let Brainy Handle Transactions

```typescript
// ✅ Recommended: Use Brainy's API (transactions automatic)
await brain.add({ data, type })
await brain.update({ id, data })
await brain.delete(id)

// ❌ Avoid: Direct storage access bypasses transactions
await brain.storage.saveNoun(noun) // No transaction protection
```

### 2. Handle Errors Gracefully

```typescript
// ✅ Recommended: Catch errors, transaction rolls back automatically
try {
 const id = await brain.add({ data, type })
 return id
} catch (error) {
 console.error('Add failed, rolled back:', error)
 // Decide how to handle (retry, log, alert user)
}
```

### 3. Validate Before Operations

```typescript
// ✅ Recommended: Validate early to avoid unnecessary rollbacks
if (!isValidVector(vector, brain.dimension)) {
 throw new Error(`Vector must have ${brain.dimension} dimensions`)
}

await brain.add({ data, type, vector })
```

### 4. Batch Related Writes with `transact()`

```typescript
// ✅ Recommended: writes that must land together go in one batch
await brain.transact([
 { op: 'add', id: orderId, type: NounType.Document, subtype: 'order', data: 'Order #1042' },
 { op: 'relate', from: customerId, to: orderId, type: VerbType.Creates, subtype: 'purchase' }
])

// ❌ Avoid: sequential single operations when partial application is unacceptable
const id = await brain.add({ ... })   // commits alone
await brain.relate({ ... })           // a crash here leaves the entity unlinked
```

### 5. Understand Atomicity Guarantees

**What Transactions GUARANTEE:**
- ✅ Atomicity within a single Brainy process
- ✅ Consistent state across all indexes
- ✅ Automatic rollback on failure
- ✅ Works with all storage adapters (filesystem, memory, custom plugin adapters)

**What Transactions DON'T Provide:**
- ❌ Two-phase commit across multiple Brainy instances
- ❌ Distributed locking across processes
- ❌ Cross-datacenter ACID guarantees

**Design:** Transactions ensure atomicity at the **write-coordinator level** inside one process. Cross-instance coordination, if you need it, lives in your service layer.

## Testing Transactions

### Unit Tests

```typescript
import { describe, it, expect } from 'vitest'
import { Brainy } from '@soulcraft/brainy'

describe('Transaction Tests', () => {
 it('should rollback on failure', async () => {
 const brain = new Brainy()
 await brain.init()

 const id1 = await brain.add({ data: { name: 'Entity 1' }, type: NounType.Thing })

 try {
 await brain.add({
 data: null as any, // Invalid - will fail
 type: NounType.Thing
 })
 } catch (e) {
 // Expected failure
 }

 // First entity should still exist (rollback didn't affect it)
 const entity1 = await brain.get(id1)
 expect(entity1).toBeTruthy()
 })
})
```

### Integration Tests

See `tests/transaction/integration/` for comprehensive integration tests covering:
- Sharding integration (`sharding-transactions.test.ts`)
- Type-aware integration (`typeaware-transactions.test.ts`)
- Distributed scenarios (`distributed-transactions.test.ts`)

The atomicity guarantees of `brain.transact()` — including crash recovery
through the real recovery path — are proven in
`tests/integration/db-mvcc.test.ts`.

## Troubleshooting

### High Rollback Rate

**Symptom:** a high share of writes throw and roll back

**Possible Causes:**
1. Invalid vector dimensions
2. Type validation errors
3. Storage write failures (disk full, network issues)
4. Index corruption

**Solutions:**
- Validate data before operations
- Check storage adapter health
- Monitor disk space and network connectivity
- Review error logs for patterns

### Slow Transaction Performance

**Symptom:** Operations take > 100ms per transaction

**Possible Causes:**
1. Large metadata objects
2. Remote storage latency
3. Many indexes enabled
4. Disk I/O bottleneck

**Solutions:**
- Optimize metadata size
- Use local caching for remote storage
- Disable unused indexes
- Use SSD storage

## Architecture Details

### Transaction Lifecycle

```
1. BEGIN
 ↓
2. ADD OPERATIONS
 - SaveNounMetadataOperation
 - SaveNounOperation
 - UpdateGraphIndexOperation
 ↓
3. EXECUTE (sequential)
 - Execute operation 1 → Success
 - Execute operation 2 → Success
 - Execute operation 3 → FAILURE
 ↓
4. ROLLBACK (reverse order)
 - Undo operation 2
 - Undo operation 1
 ↓
5. THROW ERROR
```

### Operation Types

| Operation | Description | Undo Behavior |
|-----------|-------------|---------------|
| `SaveNounMetadataOperation` | Save entity metadata | Restore previous metadata or delete if new |
| `SaveNounOperation` | Save entity data | Restore previous data or delete if new |
| `UpdateGraphIndexOperation` | Update graph index | Restore previous index state |
| `SaveVerbMetadataOperation` | Save relationship metadata | Restore previous metadata or delete if new |
| `SaveVerbOperation` | Save relationship data | Restore previous data or delete if new |

### Storage Adapter Integration

Transactions use the `StorageAdapter` interface:

```typescript
interface StorageAdapter {
 saveNounMetadata(id: string, metadata: NounMetadata): Promise<void>
 saveNoun(noun: Noun): Promise<void>
 deleteNounMetadata(id: string): Promise<void>
 deleteNoun(id: string): Promise<void>
 // ... other methods
}
```

**Key Insight:** Both shipped storage adapters (filesystem, memory) — and any custom plugin adapter — implement this interface. Transactions work with **any** storage adapter automatically.

## Additional Resources

- **Unit Tests:** `tests/transaction/Transaction.test.ts`, `tests/transaction/TransactionManager.test.ts`
- **Integration Tests:** `tests/transaction/integration/`
- **MVCC Proofs:** `tests/integration/db-mvcc.test.ts` (atomicity, CAS, crash recovery for `brain.transact()`)
- **Consistency Model:** [docs/concepts/consistency-model.md](concepts/consistency-model.md)

## Summary

Brainy's transaction system provides **production-ready atomic operations** with automatic rollback. Every single-operation write is transactional out of the box, and `brain.transact()` extends the same guarantee to multi-write batches — one atomic commit, with compare-and-swap and durable transaction metadata.

**Key Takeaways:**
- ✅ **Automatic**: No manual transaction management needed for single operations
- ✅ **Atomic**: All operations succeed or all rollback — per operation and per `transact()` batch
- ✅ **Compatible**: Works with all storage adapters and features
- ✅ **Coordinated**: Per-entity `ifRev` and whole-store `ifAtGeneration` CAS reject conflicting batches before anything is staged

Start using transactions today - they're already built into `brain.add()`, `brain.update()`, `brain.delete()`, and `brain.relate()` — and reach for `brain.transact()` whenever several writes must land together.